From own experience, the problematic parts are :

- deployment being a bit messy if you try to follow 12 factors (might be improving in 1.9)

- absence of a decent debugger

- younth of the ecosystem and size of the community (aka hard the find a maintained lib that works for a few common issues.)

- compiler being a bit slow to my taste (YMMV)

- then, of course, the elephant in the room: the absence of the kind of static type checking that facilitate refactoring ord relatively large codebases.

Nothing unbearable or permanent, but some things to keep in mind. Go try it !

And the post was very much intended to be enthusiastic about cool things, rather than nuanced and looking for problems. So I think your take-away is entirely fair.

- average performance ( like 10x slower than Java even worse for CPU intensive tasks )

- dynamic language ( this is the worst part ), working on large code base means problems ahead

- lot of features from BEAM / OTP that are not that useful and done better on modern cloud platforms ( Kubernetes for instance does a lot of similar things but better and more flexible, apply to any languages ) People like to talk about hot code update which imo is a terrible idea, you should have a proper CI/CD pipeline to do that and not rely on dangerous features like that.

- lack of ecosystem / libraries

- it's FP, I count that as a personal cons

IMO, I the fact that Elixir is a dynamic language is not that much a problem thanks to pattern matching.

Elixir's pattern matching is really powerful, allowing for things like:

    def move(_state, _x, _y, kind) when kind not in ~w(forward backwards) do
      {:error, :invalid_move_kind}
    end

    def move(state, x, y, kind) do
      get(state, x, y)
      |> do_move(state, x, y, kind)
    end

Also, there are things like structs:

    defmodule RobotVsDinosaurs.Robot do
      alias RobotVsDinosaurs.Robot

      @enforce_keys [:pos]
      defstruct name: nil, pos: nil

      def set(state, x, y, robot = %Robot{}) do
        case Matrix.elem(state, x, y) do
          %Robot{} -> {:ok, Matrix.set(state, x, y, robot)}
          _ -> {:error, :invalid_type}
        end
      end
    end

You can also do crazy things like pattern matching if a byte array have some specific characteristics, like a magic number.

So while I think it would be nice for Elixir to have types (there is Dialyzer however it is kinda a PITA to use), it is less necessary than Python/Ruby/Javascript thanks to the above.

If typing is something you really want to catch at compile time, and pattern-matching and guards aren't enough for you, you can always define Dialyzer specs and run Dialyzer during your build process. I'm currently working on an Elixir codebase that does this, but I've worked on others that don't, and I haven't seen much of a difference either way. Personally, I think strong typing is overrated.

I have come to enjoy the more self-documenting nature of static typing (in PHP of all languages). Yesterday I worked on a legacy project with no typing, and to determine what was passed to an untyped method (with no documentation) required running the interactive debugger and inspecting what was passed in different scenarios...

I've seen this stated many times as one of the big drawbacks of Beam languages. But using Distillery with Elixir, it's very easy to generate a release that's completely self contained. Just extract a tarball on a freshly installed system and you are ready to go. If your target platform differs from the one you are compiling your project on, there are various options for cross compiling too

    mix edeliver build release production && mix edeliver deploy release production

    mix deploy

Urgh, that's a disaster.

On the other hand, there's so many nice things that you can take advantage of without being a functional programming / Elixir god and every time you figure something out you think "how the heck was I able to survive before being able to do this?". I'm hardly an Elixir veteran but I'm past the Kool-aid phase. I still get amazed almost every day working on this project for 2-3 hours a day.

In practice, Elixir data structures are rarely dynamic in nature. All structs have the same fields. Virtually all lists and maps and whatnot you end up writing will have the same type for all elements. Unlike in the JS ecosystem, there is no culture of having functions with, say, optional arguments in the middle, or arguments that can be either a number or a function or a Date object or stuff like that. Unlike the Ruby ecosystem, there's no runtime monkey patching, runtime attribute generation, runtime anything generation. Everything is as robustly typed as your average C# program, but there's no way to write that down.

Or, well, there is, typespecs. And typespecs are imprecise, hard to write, borderline impossible to typecheck (dialyzer errors are shit), and generally just barely worth it because they're so hard to use.

It just feels like.. such a waste. Dynamic typing is nice if it lets you iterate faster, but that's nearly never the case in practical Elixir code, because there's so little dynamism and magic abound (this is a good thing!). At the same time, not being able to see what values a parameter can be, not being able to ever make proper IDE support, for a language and ecosystem that is so static in nature, is just a missed opportunity.

I still really like Elixir, and I agree with the GP's love letter, but I wish it had a friendlier way for me to communicate data structures to the program, future me, and my colleagues. I still hold hope that someone (future me, maybe?) will bolt types onto Elixir in the same way TypeScript bolted them onto JavaScript. It should be easier than TS because of how static-y the entire Elixir ecosystem already is.

(Sidenote: the usual argument against static types in Elixir/Erlang is that static type checking across process boundaries is hard, if not impossible. Genius Haskellers tried it and failed. My answer is that not having type checked messages would solve 95% of my itch. In the end, most lines of Elixir code is single-process, just transformations and function calls like in any other language. I'll be happy to tell the compiler what type I expect a received message to be and crash at runtime if it isn't - just like I do when I get sent a JSON payload over a REST API or IPC or whatever)

For example (taken from the docs on guards):

    def foo(term) when is_integer(term), do: term
    def foo(term) when is_float(term), do: round(term)

What's interesting about guards is you can also put expressions in there. Like `when foo > 3` and now suddenly you have a new type of integer that only accepts integers greater than 3, except it's not an explicit IntegerGT3 type you have defined somewhere.

In practice, how much worse is that vs a "real" statically typed language, other than putting guards on things is optional?

I like the convenience of deploying a golang app, just one executable.

This used to be done with Distillery[1], but I suspect its use will go down now that Elixir has native OTP releases.[2].

In either case, there's still work behind it, but it's not like deploying Ruby, Python, or Node, where you have to be pretty careful about library paths and runtime versions.

   [1]: https://github.com/bitwalker/distillery
   [2]: https://elixir-lang.org/blog/2019/06/24/elixir-v1-9-0-released/

This is crazy. What about configuration, systemd unit file / initscript, documentation, manpage, shell completion. And the ability to tell what is installed, where, and at which version.

HN forgot how to do software distribution to the hype of statically compiled languages.

The BEAM was originally designed to run software for telephone switches. It's often referred to as soft realtime because it is so responsive. It has been battle-tested for now over 3 decades.

Where the BEAM falls flat is pure number crunching but it's blazing fast with binary processing (e.g. string parsing).

A fresh phoenix project where you render a nontrivial template easily reaches sub millisecond response times, without any kind of optimization.

That being said, I agree that performance for certain IO intensive applications like web servers is very good, way better than Ruby. Subjectively Phoenix feels way, way faster than Rails in development and is much more performant in production.

Yeah, but how many companies do you know of that perform heavy computations in such large scale that their choice of programming language has a noticeable effect on the result?

I mean, it might matter for tech giants like Google, who need to squeeze every bit of processing power out of their CPUs. For almost everyone else, it will be a non-issue.

Going from Phoenix to Rocket.rs I saw a pretty dramatic speedup on even simple pages. Phoenix was faster than the Rails version of the app, but still slower than a compiled language.

From the sound of it things have gotten better (by throwing out oft-touted features), but Elixir deployments are inherently far too complicated. With Go and Rust you get a single binary and a single program to run. With Java (and other languages targeting the JVM) you get a single jar archive. With Elixir you get to bundle the whole runtime and spawn a few separate supervisor processes to run even a hello world app. Lord help you if the only maintainer for the deployment tooling takes an unannounced vacation or if the various BEAM processes stop talking to each other because the stars are misaligned (and the deployment tools are minimally tested) and you can't even run an Elixir based database migration as a result.

When I started toying around with Elixir and Phoenix I'd just come off of a Clojure kick. You can get most, if not all, of the syntactical sugar that Elixir promises with other languages that all offer infinitely simpler deployments with more mature tooling. In the case of Clojure you get whatever JVM tooling you'd like.

Make it as simple as possible, but no simpler.

First, discerning the compilation step to IR is relevant, which is why for example, Python can be compiled or interpreted. That distinction matters. Even though it is compiled to intermediate representation, it is still compiled.

Now, with regards to the intermediate representation, Java Byte Code is also not interpreted. Java Byte Code is compiled to native machine code either Just in Time by the JVM, or Ahead of Time by the SubstrateVM. Both of these make it a compiled language.

An interpreted language never gets translated into native machine code, it gets executed by a native interpreter, and that's very different. It is more akin to using a language where you read and parse texts, and based on the text, you might execute one or more things. Now if that parsing to execution branching is powerful enough to allow Turing complete behavior, you have yourself a full blown interpreted computer language. This is not what happens in the JVM. The JVM translates Java Byte Code to native machine code Just in Time, and then the native code is run.

Hotspot includes an interpreter -- your code may be interpreted until it gets hot and gets compiled.

Still, I don't think it's fair to call Java an interpreted language since the parts that are interpreted are only during warm up or slow enough not to matter.

Java on the other hand is 100% compiled (to bytecode). And byte code is machine code for a non existent machine. Actually, I think Sun had built machines that could natively execute Java Byte Code. Someone could build an interpreter for it, but I'm not aware of one.

Regardless, interpreted does not mean bad, or poor performance for your task.

It’s a technical notion.

Most people would say Python is interpreted, even if you are only running compile python bytecode on a JIT like PyPy. That’s exactly the same situation as Java.

This makes Java a compiled language, even though it is compiled to machine code for a machine that doesn't exist. Java Byte Code is an assembly language, and in fact, Sun had at one point machines that were natively using Java Byte Code instruction sets.

Now Java Byte Code is trickier. You could consider it to be interpreted or compiled or both. It is fair to say that in general it is interpreted, but it is just as fair to say it is in general compiled.

What matters most though is to understand that modern JVMs make use of a JIT compiler and optional AOT compilers.

In the latter mode, you can compile Java Byte Code to native machine executables ahead of time or pre-runtime and no JVM beyond that point is required. So you can distribute the app as a self-contained executable. In most cases though, this will be less performant in the average and peak performance, but it will speed up the worst case, such as start time.

In the former mode, the JVM will analyze runtime behavior and based on the frequency of use of various code paths, it will either compile the code to native machine code (cache the compiled code), and then run the freshly compiled code block, or it will choose to interpret the byte code directly. This allows aggressive optimization from information which is only available or easily available at runtime when performing compilation.

Actual sourced-based interpretation is _very_ slow, and pretty rare in anything that sees meaningful real world use. Ruby was, back when it was an order of magnitude slower than Python, but no other example jumps readily to mind, and Ruby went bytecode with the release of 1.9 in 2007.

> Python is an interpreted language, as opposed to a compiled one, though the distinction can be blurry because of the presence of the bytecode compiler. This means that source files can be run directly without explicitly creating an executable which is then run.

OpenJDK can't do that. You need to pre-compile the source to byte code first.

GCJ, for instance, allows AOT compilation of a Java program that requires no JVM.